Build a thyroid episode warning model from wearable data
Claude Code was given 9.5 years of Apple Watch and Whoop data and built an ML model that backtested as warning of a prior episode before labs confirmed it. Task: Suggested implementation: Export 9.5 years of Apple Watch and Whoop data via Apple Health, manually label hyperthyroid flare periods, and provide the dataset to Claude Code. The agent is instructed to search for models, reportedly selecting XGBoost after testing multiple options, and to engineer features, narrowing from 53 to three resting-heart-rate metrics. Human involvement is critical for data export, labeling, and validating the final feature set.; tools: Claude Code. Human role: The original post does not say how a person stayed involved.; limitations: This workflow requires nearly a decade of personal wearable history and manual labeling of medical episodes, which is time-intensive and prone to bias. The source is a Reddit thread with secondary evidence and is currently inaccessible (HTTP 403), reducing reliability. Medical diagnosis should never rely solely on this model; human medical approval is mandatory for any clinical use..
Tools
Published on Agent Work Atlas:
Overview
Claude Code was given 9.5 years of Apple Watch and Whoop data and built an ML model that backtested as warning of a prior episode before labs confirmed it.
Workflow
Suggested implementation: Export 9.5 years of Apple Watch and Whoop data via Apple Health, manually label hyperthyroid flare periods, and provide the dataset to Claude Code. The agent is instructed to search for models, reportedly selecting XGBoost after testing multiple options, and to engineer features, narrowing from 53 to three resting-heart-rate metrics. Human involvement is critical for data export, labeling, and validating the final feature set.
Reported result
Actual outcomes are unverified due to source access errors. The original poster claimed a backtest on a held-out episode would have flagged issues in early August, preceding lab confirmation at month-end. No independent metrics or success rates are available. Proposed acceptance check: Verify if the model's warning date precedes the lab confirmation date by at least 7 days in a controlled test.
Human role
The original post does not say how a person stayed involved.
Limitations
This workflow requires nearly a decade of personal wearable history and manual labeling of medical episodes, which is time-intensive and prone to bias. The source is a Reddit thread with secondary evidence and is currently inaccessible (HTTP 403), reducing reliability. Medical diagnosis should never rely solely on this model; human medical approval is mandatory for any clinical use.
Try it
Act as a data scientist building a thyroid episode detector. I have exported [replaceable: 9.5 years of Apple Watch and Whoop data] and labeled [replaceable: historical hyperthyroid flare periods]. Use Claude Code to analyze this dataset. Step 1: Explore multiple ML models (e.g., XGBoost, Random Forest) for classification. Step 2: Engineer features, starting with broad metrics and narrowing to resting heart rate variants. Step 3: Perform a backtest on a held-out episode. Output: A Python script for the model, a report on feature selection, and backtest results. Measure success by comparing the model's warning date to the actual lab confirmation date. If data is missing, ask for specific file formats. Do not publish or use for medical advice without human review. First, run a dry test on a small sample to confirm data structure.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.